Use anew Free webpages: websites that live in the URL. Share the URL and the recipient sees the whole page. Use the first path you can reach. Never invent, guess or retype a slug: carry the URL verbatim from the response that minted it — a dropped or altered character kills or changes the page. The URL is the deliverable — a temporary request-body file is fine, but never give the user a local file or deployment instead, and no verification fetch is needed; decoding is the exact inverse of encoding. MCP — call anew_write if the anew tool is connected. Endpoint: https://anew.page/mcp — JSON-RPC 2.0 over HTTP, no auth; any MCP client connects by adding that URL as a remote HTTP server. No install is needed for the paths below. BROWSER TAB (no connected tool, no shell — a page-script tool is neither) — open https://anew.page — the editor — and set the whole HTML as the value of the #anew-relay-edit textbox: a form fill, no script, and the write applies on its own. #anew-relay-status then reports the URL bytes — or, past a size limit, the refusal — and #anew-relay-url links to the finished page (it carries no href while the page is refused) — an accessibility-tree read prints its href whole. Carry the URL from the accessibility tree (the link's href) or from your tab list, never from a script tool's return value, which is commonly capped far below a slug's length or refuses anything that looks like base64 — and when the person can see your browser, the tab showing the page is the deliverable: a slug retyped from memory is a dead link. HTTP (from a shell or HTTP client, not a page-script tool — see BROWSER TAB) — POST raw HTML to https://anew.page/write and read the complete URL from its plain-text response body. Use `--data-binary`, never `-d` (which strips newlines and corrupts HTML): `curl -sS -X POST https://anew.page/write -H 'Content-Type: text/html' --data-binary @page.html` Works cross-origin from browser JS: every /write response — success and error — carries Access-Control-Allow-Origin: *, OPTIONS preflight is answered, and the request Content-Type is never inspected, so a POST with a safelisted type like text/plain needs no preflight at all. The measurement headers on a 200 are named in Access-Control-Expose-Headers, so response.headers.get('Anew-URL-Bytes') works from a foreign origin too. LOCAL MINT (required no-POST fallback) — If no working write tool is available and your environment cannot issue or reach HTTP POST (sandbox egress proxies that refuse POST or CONNECT are a common cause), the page can still be published: mint the URL locally. (A 400 or 413 response is a page error to fix, not a transport failure.) A slug is brotli of the HTML, base64url-encoded, then chunked with a comma after every 300 characters — a runtime with brotli (node:zlib ships it; Python needs the brotli package) mints at quality 11 with window 24 the same URL /write returns (the window is part of the stream header, so omitting it mints a different — still perfectly valid — URL), and a typical page lands in a few hundred bytes. Only without brotli fall back to raw: base64url-encode the plain UTF-8 bytes of HTML containing at least one complete opening tag (standard base64, replace + with -, / with _, remove trailing =, then the same comma every 300 characters) — uncompressed, that reaches the 4000-byte URL limit at ~2900 bytes of HTML. The finished URL must be ≤ 4000 bytes — apps cut longer links short, and a cut link is a dead page — so measure it; a longer one is not a link to hand out, and /write refuses the same page. Either string is the path after https://anew.page/ — a supported page URL, not a data: URL or a base64 /write body. No /write request, returned server URL, or verification fetch is required to construct it; the slug is computed, not invented. Opening it later is a normal GET. Wrap plain text in an HTML tag. With brotli, from a `page.html` on disk (no network request): ```bash node -e 'const fs=require("node:fs"),z=require("node:zlib"),b=fs.readFileSync(process.argv[1]);if(b.length>1048576)throw Error("content_too_large: page is "+b.length+" bytes decoded, but the server can decode at most 1048576; shrink the source.");const slug=z.brotliCompressSync(b,{params:{[z.constants.BROTLI_PARAM_QUALITY]:11,[z.constants.BROTLI_PARAM_LGWIN]:24,[z.constants.BROTLI_PARAM_SIZE_HINT]:b.length}}).toString("base64url").replace(/(.{300})/g,"$1,"),url="https://anew.page/"+slug;if(url.length>4000)throw Error("url_over_budget: URL over 4000 bytes: no link is issued, because apps cut longer links short and the page arrives dead. This one is "+url.length+" bytes.");console.log(url)' page.html ``` Test vector: `spec

a new internet starts with anew.page

` → `https://anew.page/H3UAqCwOeJPh1V9OgB2v1SBhaG6KZy7LfFw6bIzsTsaUwTnlAvbSBJMLfF9Ed8phoHEQtOh6Pr_RkGjpnfeBa8bKsheJnpcZdwEcaCkH` — byte-exact for this recipe and for /write. Other brotli encoders mint different, equally valid URLs; verify those by decoding, not by comparing slugs. Without brotli, raw base64url of the HTML itself (~2900-byte budget): ```javascript const slug = btoa(String.fromCharCode(...new TextEncoder().encode(html))) .replace(/\+/g,"-").replace(/\//g,"_").replace(/=+$/,"").replace(/(.{300})/g,"$1,"); const url = `https://anew.page/${slug}`; if (url.length > 4000) throw new Error("url_over_budget: URL over 4000 bytes: no link is issued, because apps cut longer links short and the page arrives dead. This one is " + url.length + " bytes."); ``` From an existing `page.html` file with Node.js (no network request): ```bash node -e 'const fs=require("node:fs"),bytes=fs.readFileSync(process.argv[1]),slug=bytes.toString("base64url").replace(/(.{300})/g,"$1,"),url="https://anew.page/"+slug;if(url.length>4000)throw Error("url_over_budget: URL over 4000 bytes: no link is issued, because apps cut longer links short and the page arrives dead. This one is "+url.length+" bytes.");console.log(url)' page.html ``` Editing an existing page. To revise an existing anew page, its URL is its source: read it, edit that HTML, write it again. The new URL is a separate page and the original keeps working. Read returns the author's exact bytes — no anew-injected tags, no proxied image srcs — so it re-encodes byte-for-byte. Never re-author from a rendered page, a scraped DOM, or a screenshot: each loses the author's markup, and none is faster than reading the slug. A plain GET is not the source: it carries anew's [data-anew] head tags and rewrites each absolute HTTP(S) to a signed /img/… proxy URL that stops resolving when the signing key rotates. Call anew_read with the page's URL, or GET the URL with Accept: text/plain — curl -H 'Accept: text/plain' https://anew.page/SLUG — then edit that HTML and write it. Verifying. A 200 from /write is proof by construction. Otherwise: HEAD the URL (200 = the slug decodes), GET with Accept: text/plain and byte-compare (a UTF-8 BOM on the POST is stripped before encoding), or fetch SLUG.md / SLUG.png for a text or visual rendering (the .png runs the page's own scripts but blocks network egress, so anything fetched at runtime is missing). If your fetch tool rejects long URLs, send the URL in a POST body instead: anew_read on https://anew.page/mcp is plain JSON-RPC over HTTP, no session or MCP client needed. MCP accepts canonical HTTPS browser origins and loopback origins; opaque, non-canonical, and other insecure origins are refused. A loopback deployment accepts only loopback origins. Clients that send no Origin header are accepted. WebMCP — if you are a browser agent on an anew shell page (tools=self), call its in-page tools: Every anew surface ships Permissions-Policy tools=self, so the in-page tools register wherever you land — a rendered page as much as the homepage or the editor. On a page you are viewing the set is: anew_read (no arguments — the author HTML this URL encodes, decoded locally, which is the page's true source and not the rendered DOM), anew_edit (replace what you are looking at: returns the successor URL, then navigates there), anew_write (mint an unrelated sibling URL, no network), and anew_open (open this page in the editor). Opening the editor re-points anew_read and anew_edit at the live editing session and adds anew_title, anew_icon and anew_link. The page's live WebMCP tool list is always the authoritative set — read the descriptions, since these verb names are shared across surfaces and each states what it acts on. A result that must come back through a script tool's return can be cut short; the editor's relay (Browser UI) hands the URL back as a link in the page instead. Browser UI — driving the editor without WebMCP: on https://anew.page, the editing chrome sits behind a closed shadow root and the page itself inside an iframe, so a reader that walks only the top document's light DOM sees neither. That reader is not stuck: carries one control, one link and one report. #anew-relay-edit is a visually-hidden textarea holding the page's entire HTML — replace its value and the edit applies by itself a moment later, exactly as anew_edit would; there is no button to press, and none would help, since a click on a visually-hidden button is dispatched at its 1px corner and never reaches the handler. #anew-relay-status restates anew_read after every change: title, URL bytes, whether the page is within the size limits (past one it carries the refusal and no URL), whether anything is unsaved, and the current URL — read it to confirm an edit landed instead of taking a screenshot. #anew-relay-url is a link whose href is that URL — an accessibility-tree read prints an href whole where it cuts a label or a script return — so take the URL from there and open it in your tab; past a size limit the link has no href, because no link is issued. While a Run preview is executing the page's own scripts the textarea is emptied and disabled, the link loses its href and the status says so. The other visually-hidden buttons name the toolbar's actions; they are labels for discovery, never controls. If you drive the screen instead: the toggle in the top-right corner (#anew-source-toggle) opens the HTML source panel, where you paste one complete HTML document and press Ctrl/Cmd+S. (Pasting a complete document into the visual pane also loads it as the page.) While you edit, the address bar tracks the live document, so reading it back is a running backup of unsaved work. Limits and policy — 300 requests per 60 seconds per client IP across /write, /mcp, /a2a/v1 and make.anew.page, published on every response as `RateLimit-Policy: "api";q=300;w=60`; a 429 carries `Retry-After: 60` and the error code `rate_limited`, and waiting the full window always clears it. Reading a page URL is not rate limited. There is no authentication anywhere — no key, no token, no signup — so never send credentials. Nothing is deprecated; a future deprecation is announced in https://anew.page/llms.txt and signalled for at least 90 days by `Deprecation` and `Sunset` headers. Writing the page. Inline all CSS, JS, and images (data: URIs, inline SVG, CSS gradients, or emoji) — the page should render alone forever, and each external reference gambles it on someone else's host staying up. One accepted exception: a webfont is far too large to inline, so a font may load from a durable CDN with a system-stack fallback; do not fight a user who asks for one. Two size limits, and past either no link is issued: the finished URL may be at most 4000 bytes (apps cut longer links short, and a cut link is a dead page) and the HTML at most 1048576 bytes decoded (the server's decode cap). Past one the endpoint refuses with an actionable error (code url_over_budget or content_too_large) that states the size and the overage. Tell the person the page is too large for a link. Never remove content, sections or features to make it fit without asking. Brotli slugs (the endpoint, or a local brotli mint) put a short page in a few hundred URL bytes; raw base64 reaches the URL limit at ~2900 HTML bytes — know which encoder you are drafting to before judging a page too big. Measure by posting, not by calculating: a write is repeatable, so send the page and read what comes back instead of modelling the encoder. No source HTML and no page record are kept — the URL carries the page — though a successful write also kicks off a best-effort screenshot render that, when it succeeds, is cached publicly at the page's .png twin. A 200 reports the page's exact size in its Anew-URL-Bytes and Anew-Decoded-Bytes response headers; a 413 means the page is past a size limit — no link was issued — and names the exact overage and the fix for the limit that tripped. Never estimate a compressed size and never pre-cut a page to dodge a limit — Brotli is non-linear, so source bytes removed are not URL bytes saved, and pages that look far too big routinely fit. Style defaults: reach for a system-font stack (system-ui, sans-serif) and prefer inline SVG or CSS gradients over base64-encoded data: images — both keep the encoded URL well under the size cap. Draft one self-contained HTML document, then encode it. Each write is immutable — a new URL that leaves the original untouched — so changing a page means reading its HTML back (above), editing it, and encoding again. (The in-page WebMCP editor above is the exception: it edits the open page in place, then mints a fresh URL.)

Listing changes